Copilot in Word: Drafting, Editing, and Summarising at Scale
Word is the application most office workers spent the early 2020s pasting in and out of ChatGPT for. Copilot in Word collapses that round trip into a side panel. Drafting from a brief, rewriting paragraphs in different tones, summarising long contracts, comparing versions, building citation lists — the workflows that consultants once charged for now run inside the document. The output quality is uneven, the time savings are real, and the trick is knowing which moves Copilot is good at and which it still gets wrong.
Table of contents
- Drafting a document from a brief
- Rewrite Anywhere — the underused feature
- Summarising long documents
- Comparing two documents
- Citation and reference handling
- Style consistency across long drafts
- Frequently asked questions
- The bottom line
Drafting a document from a brief
The headline feature: hand Copilot a one-paragraph brief and a reference document, and it produces a structured first draft. The cleanest prompt pattern is three parts. Tell Copilot what you want, who it is for, and what tone. "Draft a 1,500-word client briefing on the new VAT changes for our SME customers in plain English." That produces something usable in 15 seconds.
Where the magic actually lives is the "Reference" field. Click the paperclip in the Copilot draft panel and attach up to three documents from OneDrive — past briefings, the company tone-of-voice guide, source data. Copilot uses these as grounding. The output picks up your house structure, your typical paragraph length, and your default disclaimers. The draft is no longer generic LLM prose; it is your prose-shape filled with new content.
Two warnings. First, Copilot's drafts run long by default. A "1,500-word" instruction usually returns 1,800 to 2,000 words because the model overshoots. Either ask for the shorter target deliberately or expect a cut after the first pass. Second, Copilot will fabricate statistics if you ask it to "include supporting data" without providing source documents. Treat any number that didn't come from your reference attachment as suspect — the same rule that applies to prompt engineering generally.
Rewrite Anywhere — the underused feature
Rewrite Anywhere arrived in late 2024 and remains the single most underused Copilot feature. Highlight any paragraph in any document, and Copilot offers three rewrites in a side panel — typically labelled shorter, punchier, and more formal. The rewrite happens in place; you accept the version you want and move on.
The productivity gain is in the friction it removes. Before this feature, the rewrite workflow was: select text, copy to ChatGPT, ask for a rewrite, paste back, fix the formatting Word lost. Five steps. Now it is one click and one decision. On a 4,000-word document where you might rewrite a dozen paragraphs, the time saved adds up to 20 to 30 minutes.
The prompt is implicit but you can override it. Click the small "tone" dropdown in the rewrite panel for options: professional, casual, encouraging, neutral, witty. The witty setting is rarely useful in business writing but occasionally produces a sharper version of a flat sentence. For specialised house-style work — legal, medical, regulatory — the rewrites are competent on neutral and formal but should always be edited; technical accuracy is not what they optimise for.
| Word feature | What it does | Time saved per doc | Reliability |
|---|---|---|---|
| Draft from brief + reference | Generates first draft from prompt and attached docs | 30-60 min | High on structured docs, low on creative |
| Rewrite Anywhere | In-place rewrites in three tones | 15-30 min on long docs | High |
| Summarise this document | Structured outline of contents | 20-45 min on long docs | High on text-heavy, lower on tables/figures |
| Compare two versions | Plain-English diff highlighting substantive changes | 20-30 min vs Track Changes | Medium |
| Generate citations | Reformats and inserts source list | 10-15 min | Medium — verify each entry |
| Style consistency check | Flags tone and terminology drift | 10-20 min | Variable |
Summarising long documents
Drop a 40-page contract into Word, hit the Summarise button at the top of the Copilot panel, and you get a structured brief: parties, key obligations, key dates, exceptions, anything the model flagged as unusual. The output is honest about uncertainty — it will tag clauses as "review" rather than asserting it understands them.
The 2025 update added scoped summarisation. Instead of summarising the whole document, you can summarise a specific section, a specific topic across the document, or a comparison between two sections. The legal team at a UK insurer reported in early 2025 that Copilot-led first-pass review of vendor contracts dropped from 90 minutes per contract to 25 — not because Copilot replaced lawyers, but because it routed attention to the parts that actually warranted lawyer time.
Where summarisation degrades: documents that are mostly tables, mostly figures, or mostly footnotes. Copilot reads the body text well; it does not read complex visual layouts well. For technical specs, regulatory filings, and financial reports with dense tabular data, expect the summary to mention the prose around the tables but not faithfully describe what's inside them.
The summarisation feature most people miss: cross-document summarisation in Copilot chat. Outside Word, in the M365 Copilot chat surface, you can ask "summarise the last three Acme Corp contracts" and Copilot will pull the documents from SharePoint, summarise each, and produce a comparison. That is genuinely powerful for anyone working with families of related documents — sales sees three competing customer proposals at a glance, legal sees how a clause has evolved across vendor contracts, finance sees how forecasts shifted across quarterly board packs. The chat surface is the place to do this; the in-Word summarise button only handles one document at a time.
Comparing two documents
"Compare these two contract drafts and tell me what changed materially" is one of those prompts that sounds simple and hides genuine value. Track Changes shows you every change including formatting and whitespace. Copilot triages — it tells you which changes affect liability, scope, payment, or termination, and which are cosmetic.
Open the two documents, invoke Copilot, and ask "Compare with the version at [link or attachment]." Copilot returns a categorised diff in plain English: substantive additions, substantive deletions, and stylistic changes. On a 30-page redline, this is the difference between an hour of careful reading and a 10-minute briefing.
The catch: comparison quality depends on the documents being similar enough. Comparing draft 4 against draft 3 works well; comparing draft 4 against the original RFP response does not, because the structure has shifted too far. Use Copilot comparison for evolutionary changes, Track Changes for revolutionary ones.
The comparison feature has one underrated extension: ask Copilot what changes were not made. "Compare these two drafts and tell me what feedback from the email thread was not incorporated." Copilot pulls from the email context, identifies the changes that should have happened, and flags the ones that didn't. This is the closest a tool gets to playing the role of a thorough reviewer — catching what isn't in the document but should be. For high-stakes drafts where missed feedback is a real risk, this prompt earns its keep on the first use.
Citation and reference handling
Citation handling in Word + Copilot is genuinely improved but requires a mental adjustment. The classic Word Citations & Bibliography ribbon still works for managing a reference list. What Copilot adds is the ability to ask "format the references in this document in Harvard style" or "convert all citations from APA to Chicago" and have the document update.
The serious limitation: Copilot does not validate that your references exist. If you ask it to "add citations to support the claim that prompt engineering is the top in-demand skill in 2026," it will produce plausibly formatted citations to journals and articles that may not exist. This is the most dangerous failure mode of Copilot in Word. For academic, legal, or regulated writing, every Copilot-generated reference must be verified against the source. Treat the formatting work as a time-saver, the existence claim as a liability.
The workflow that does work for citation-heavy writing: research first in Copilot chat with web access enabled, capture the actual sources, then drop the source list into Word and ask Copilot to format. The mental model is "the model produces text, Bing produces facts" — keep them separate and Copilot becomes useful again. Mix them and you produce confidently fabricated bibliographies, which are worse than no bibliography at all.
Style consistency across long drafts
For documents over 5,000 words, style drift is real — early sections are crisp, late sections are bloated, and the whole thing reads like it was written over three weeks (because it was). Copilot's "Improve style" workflow runs through the document flagging tone shifts, terminology inconsistencies (does the document use "client" or "customer"?), and passive-voice density.
The output is a sidebar of suggestions, not auto-applied changes. This is correct — house-style decisions should be human. But it surfaces problems you would otherwise miss until your editor caught them. For long-form work, this is the single highest-leverage Copilot use after summarisation. The complementary perspective on tone-tuning lives in our AI writing assistants comparison.
The style-consistency workflow that works best on documents over 10,000 words is iterative: run the consistency check on the first half, address the flagged issues, then run it on the second half. Doing both halves in one pass produces too many suggestions to act on. Splitting the document into review chunks keeps each pass actionable. Most internal style guides at large enterprises ended up restructuring their long-form review processes around this pattern in 2025 — the consistency-check pass replaces a fraction of what an editor used to catch in line edit, and frees the editor's time for the higher-value structural feedback only humans do well.
Frequently asked questions
Can Copilot in Word write a document from scratch?
Yes, but the output quality depends entirely on what you give it. A one-line prompt produces generic content. A detailed brief plus 1-3 reference documents from your OneDrive produces a competent first draft in your house style. The reference attachment is the difference between Copilot output you'll throw away and output you'll edit and keep. Always attach references when drafting anything that should sound like your organisation.
Will Copilot keep my document private?
Yes. Microsoft 365 Copilot processes prompts and document content under the same enterprise data protection terms as the rest of the M365 stack — your data stays in your tenant region, is not used to train OpenAI's foundation models, and is not shared cross-tenant. Sensitivity labels are honoured: a Confidential document cannot have its content summarised into a Public destination.
Why does Copilot's draft not match my house style?
Without a reference document, Copilot defaults to a generic professional tone. Attach 1-3 examples of your existing house writing — a previous report, an internal style guide, a typical briefing — and the draft picks up structure, vocabulary, and paragraph rhythm. The improvement is dramatic. If you draft regularly, save a "house style sample" Word doc in OneDrive specifically to attach as a Copilot reference.
Can Copilot generate citations that actually exist?
Sometimes yes, often no. Copilot is generating plausible citation formats based on its training data, which may include real sources or may produce fabricated ones. For any document that depends on citation accuracy — academic, legal, regulatory, journalistic — every Copilot-suggested reference must be independently verified. Use Copilot for citation formatting (APA to Chicago, etc.) which is reliable; never use it as a research substitute.
Does Copilot work on documents stored locally?
For full functionality (drafting from references, summarising attached docs), the document needs to be in OneDrive or SharePoint. Copilot can still work on a local document — drafting, rewriting, summarising — but cross-document features and grounding against your tenant's other content require cloud storage. Most organisations save Word docs to OneDrive by default, so this rarely matters in practice.
The bottom line
Copilot in Word is at its best when you give it specific raw material: a brief, a reference document, a paragraph to rewrite, a contract to summarise. It is at its worst on open-ended creative tasks and on anything requiring factual research. Treat it as a fast junior writer with a perfect memory of your past documents and a tendency to invent citations.
The four moves with the best ROI: drafting from brief plus reference attachment, Rewrite Anywhere on long-form work, summarisation of long contracts and reports, and document comparison for redlines. Together they save most knowledge workers an hour a day on document-heavy work.
The habit that separates power users from occasional users is reaching for Copilot first, not last. The instinct to draft a paragraph manually and then ask Copilot to "polish" it produces worse output and takes longer than asking Copilot to draft from the brief and then editing. The same applies to summarisation — opening a long document and trying to skim before invoking Copilot is slower than handing it to Copilot first and then reading the summary critically. Within a fortnight of reversing this instinct, most users find their typing has gone down and their editing has gone up. That is the productivity shape Copilot in Word is best at producing.
For the wider Microsoft Copilot picture, see our complete training guide, or browse all Microsoft Copilot articles.
Last updated: January 2026